Whether an AI art generator learning from your online photos counts as stealing depends on what you mean by "stealing" — legally, it usually isn't theft, but it may still be copyright infringement, and courts have not settled that question yet.
If you post a photo publicly, an AI company can generally scrape it for training without breaking into anything, so there's no "theft" in the criminal sense. But copyright law gives you the exclusive right to control certain uses of your work, and whether training a model on your photo counts as one of those controlled uses is exactly what's being fought over right now. So the honest answer is: not stealing in the ordinary sense, possibly infringing in the legal sense, and unresolved in practice.
What "stealing" actually means here
When people say an AI art generator "stole" their photo, they're usually describing two different things at once. The first is copying — the model ingested their image without asking. The second is output — the model can now produce images that resemble their style or, in some cases, their actual photo.
Those are separate legal issues. Copying for training may fall under fair use or an equivalent doctrine in some places; producing a near-duplicate output is much harder to defend. A photographer whose mountain shot ends up reproduced almost exactly by a generator has a stronger claim than a photographer whose general aesthetic influenced the model's style.
Style itself is generally not protected by copyright — only specific expression is. That distinction matters, and it's why two people looking at the same AI output can disagree so strongly about whether a theft occurred.
The mechanism is worth understanding because it explains why this is so murky. Image generators like Midjourney are trained on enormous collections of images, and according to our AI tool database, Midjourney is developed by Midjourney Inc. and sits in the image category with an editorial rating of 4.8/5.
The database also notes its current feature set includes V7 with Draft Mode, Omni Reference, Personalization v2, Niji 7 anime, and a V1 Video Model. None of that tells you where the training images came from — and that's the point. Most major generators have not fully disclosed their training data, so an artist often cannot confirm whether their work was included.
You can sometimes check by prompting the model with your name or a distinctive title and seeing whether it produces something recognizably yours, but that's an imperfect test, not proof. Some companies offer opt-out mechanisms or honor "do not train" signals in metadata, but these are voluntary and inconsistent. There is no universal switch that removes your photos from every model.
A concrete example
Say you're a wedding photographer. You post a gallery of a client's ceremony on your public portfolio site. Six months later, a friend sends you an AI-generated image that looks strikingly like one of your shots — same composition, same lighting, same unusual cake design.
You didn't license anything. What can you actually do? First, you'd need evidence the model was trained on your specific image, which usually means the company disclosing its dataset or your image appearing in a leaked dataset index.
Second, you'd need to argue that the output is substantially similar to your protected expression, not just your general style. Third, you'd need to file in a jurisdiction whose courts have addressed AI training — and as of now, the major cases are still working through the system.
In practice, most photographers in this position find that the cost of a lawsuit outweighs the likely recovery, which is why so many disputes get resolved by public pressure or platform policy changes rather than court rulings.
Where this advice breaks down
The biggest limit is that the law here is genuinely unsettled and varies by country. The United States has a fair use doctrine that some courts have applied to training; the European Union has a text and data mining exception with its own conditions; Japan has a notably permissive approach to training data.
So the same act — training on your public photo — can be lawful in one place and contested in another. Second, "public" does not mean "free to use." Posting a photo publicly does not waive your copyright.
But it does make scraping easier and harder to detect. Third, opt-out tools are not retroactive: if your image was already in a dataset, opting out now may not remove it from a model that's already trained. Fourth, this answer is general information, not legal advice — if you're considering a claim, talk to an IP lawyer in your jurisdiction.
And finally, the practical takeaway for most creators is defensive: watermark your work, embed copyright metadata, limit full-resolution uploads, and check whether platforms you use have AI-training clauses in their terms. None of that is a guarantee, but it changes the odds.